In short. The 2023 GEO paper defines content interventions and answer visibility metrics whose reported effects require evaluation in the engine setting being tested.
Definition: The 2023 GEO paper is GEO: Generative Engine Optimization, a study by Pranjal Aggarwal and colleagues that introduces a framework for changing web content and measuring its visibility in answers produced from retrieved sources. GEO stands for generative engine optimization.
Key facts
- The arXiv record dates the first submission to 16 November 2023 and records acceptance to KDD 2024.
- The paper's experimental setup uses the top five Google sources for each query and the gpt3.5-turbo model to generate answers.
- Bajemon and Rochet's September 2026 follow-up paper re-tests the strongest content levers across ten engine families and reports no effect in the published direction under its controlled protocol.
Which version does “2023” refer to?
The submission history identifies the original preprint as version v1. It records version v3 on 28 June 2024, following an intervening revision. The author list comprises Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande.
The following identifiers distinguish the initial submission from the revised text used for the experimental details below:
arXiv:2311.09735v1
arXiv:2311.09735v3
A reference to the “2023 paper” therefore identifies the submission year; the arXiv version identifies the text being discussed.
What does the experiment measure?
The authors' methodology defines visibility through the amount and placement of answer text attributed to a source. Its Word Count metric divides the words in sentences citing a source by the answer's total word count. When a sentence cites several sources, its word count is shared equally among them.
Position-Adjusted Word Count adds a declining weight for sentences appearing later in the answer. Subjective Impression uses model-based evaluation of factors including relevance, influence and the perceived likelihood of clicking a citation.
GEO-bench, the study's evaluation dataset, contains 10,000 queries. Its training, validation and test splits contain 8,000, 1,000 and 1,000 queries respectively. For each query, the experiment randomly selects one source for modification and applies the different methods separately to that same source. The baseline uses unmodified source content.
How should the reported gains be read?
The paper's results report relative improvements of 30% to 40% in Position-Adjusted Word Count for Cite Sources, Quotation Addition and Statistics Addition. These methods add references, quotations or quantitative information to the source text. The percentages describe changes in the defined answer visibility metric; the experiment compares edited and unedited content under its specified engine setup.
Bajemon and Rochet's later study reports that these levers fail to transfer in the published direction across ten tested engine families. Its full text describes a controlled comparison against neutral additions of the same length, using five curated sources and a fixed citation-format prompt. For website decisions, the two studies support measuring an edit against a baseline on the engine and queries being evaluated, while keeping the metric and experimental setting explicit.
Related terms
- Generative engine optimization (GEO)
- AI citation
- AI visibility
- The GEO score paper (arXiv 2609.07559)
FAQ
Does the paper measure website traffic?
The evaluation section measures answer visibility through Position-Adjusted Word Count and Subjective Impression. The clicking component is a model-assessed likelihood within Subjective Impression.
Can I use the reported percentages as a current citation forecast?
The September 2026 follow-up finds no positive transfer of the tested levers under its protocol. You can use the original paper's metric definitions to specify a measurement, while evaluating results in the setting you actually test.
Sources
-
arXiv, "GEO: Generative Engine Optimization", submitted 16 November 2023, revised 28 June 2024
-
Aggarwal et al., "GEO: Generative Engine Optimization", revised 28 June 2024, full text
-
Bajemon and Rochet, "Scoring Without the Engine", 7 September 2026, full text
Check how a page reads to answer engines. Open the GEO score analyzer